libm implementations are a little slower sometimes, but imo it is well worth the cost to achieve real cross-platform consistency, especially in scientific computing

I recommend using libm for any functions with that label unless you know they'll get compiled to a single instruction on your target platform (like sqrt() for example)

If you load a C module that uses -ffast-math, or if libc gets hooked and modified, it can change floating-point calculation behavior in your otherwise-separate program. I think there are also some function implementations that don't write all the bits of the float in some cases, so you get some junk